From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../sve_hybrid_u8u32_mmla_6x4VL/generic.cpp | 1809 ++++++++++---------- 1 file changed, 904 insertions(+), 905 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp index 59f33289b4..9269576d90 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_mmla_6x4VL/generic.cpp @@ -100,16 +100,16 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "incw x20\n" "whilelt p1.s, x20, x11\n" "tbz %x[flags], #0, 3f\n" - "ld1w { z9.s }, p4/Z, [x9]\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z19.s }, p4/Z, [x9]\n" + "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" + "zip1 z9.d, z18.d, z13.d\n" + "zip2 z13.d, z18.d, z13.d\n" + "zip1 z10.d, z17.d, z14.d\n" + "zip2 z14.d, z17.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" "b 4f\n" @@ -127,11 +127,11 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "5:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 6f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" "cbnz x28, 7f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -143,86 +143,86 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "ble 9f\n" "8:" // Height 1: Multiply loop: Main loop head "whilelt p0.b, XZR, x27\n" - "ld1rqb { z1.b }, p0/Z, [x26]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + "ld1rqb { z20.b }, p0/Z, [x26]\n" + "trn1 z18.d, z20.d, z19.d\n" + "ld1b { z17.b }, p5/Z, [x10]\n" + "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d19a48 // ummla z8.s, z18.b, z17.b\n" + ".inst 0x45d09a4c // ummla z12.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d19a49 // ummla z9.s, z18.b, z17.b\n" + ".inst 0x45d09a4d // ummla z13.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45d19a4a // ummla z10.s, z18.b, z17.b\n" + ".inst 0x45d09a4e // ummla z14.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - "trn2 z1.d, z1.d, z2.d\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-8, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #-7, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-6, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #-5, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-4, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #-3, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #-1, MUL VL]\n" + "trn2 z20.d, z20.d, z19.d\n" + ".inst 0x45d19a4b // ummla z11.s, z18.b, z17.b\n" + ".inst 0x45d09a4f // ummla z15.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n" + ".inst 0x45d19a88 // ummla z8.s, z20.b, z17.b\n" + ".inst 0x45d09a8c // ummla z12.s, z20.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #-5, MUL VL]\n" + ".inst 0x45d19a89 // ummla z9.s, z20.b, z17.b\n" + ".inst 0x45d09a8d // ummla z13.s, z20.b, z16.b\n" + "ld1b { z16.b }, p5/Z, [x10, #-4, MUL VL]\n" + "ld1b { z7.b }, p5/Z, [x10, #-3, MUL VL]\n" + ".inst 0x45d09a8a // ummla z10.s, z20.b, z16.b\n" + ".inst 0x45c79a8e // ummla z14.s, z20.b, z7.b\n" + "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n" "sub x27, x27, #0x10\n" "cmp x27, #0x10\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" + ".inst 0x45d19a8b // ummla z11.s, z20.b, z17.b\n" + ".inst 0x45d09a8f // ummla z15.s, z20.b, z16.b\n" "add x26, x26, #0x10\n" "bgt 8b\n" "9:" // Height 1: Multiply loop: Single iteration only "whilelt p0.b, XZR, x27\n" "ld1rqb { z1.b }, p0/Z, [x26]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + "trn1 z18.d, z1.d, z19.d\n" + "ld1b { z17.b }, p5/Z, [x10]\n" + "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d19a48 // ummla z8.s, z18.b, z17.b\n" + ".inst 0x45d09a4c // ummla z12.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d19a49 // ummla z9.s, z18.b, z17.b\n" + ".inst 0x45d09a4d // ummla z13.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45d19a4a // ummla z10.s, z18.b, z17.b\n" + ".inst 0x45d09a4e // ummla z14.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n" "subs x27, x27, #0x8\n" - "trn2 z1.d, z1.d, z2.d\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" + "trn2 z1.d, z1.d, z19.d\n" + ".inst 0x45d19a4b // ummla z11.s, z18.b, z17.b\n" + ".inst 0x45d09a4f // ummla z15.s, z18.b, z16.b\n" "addvl x10, x10, #8\n" "ble 10f\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" + "ld1b { z17.b }, p5/Z, [x10]\n" + "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d19828 // ummla z8.s, z1.b, z17.b\n" + ".inst 0x45d0982c // ummla z12.s, z1.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d19829 // ummla z9.s, z1.b, z17.b\n" + ".inst 0x45d0982d // ummla z13.s, z1.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45d1982a // ummla z10.s, z1.b, z17.b\n" + ".inst 0x45d0982e // ummla z14.s, z1.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45d1982b // ummla z11.s, z1.b, z17.b\n" + ".inst 0x45d0982f // ummla z15.s, z1.b, z16.b\n" "addvl x10, x10, #8\n" "10:" // Height 1: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -258,21 +258,21 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "whilelt p1.s, x20, x11\n" "tbz %x[flags], #0, 14f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "add x20, x9, x20, LSL #2\n" + "ld1w { z18.s }, p4/Z, [x9]\n" + "ld1w { z2.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z17.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x24]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" + "ld1w { z12.s }, p4/Z, [x20]\n" + "zip1 z8.d, z18.d, z12.d\n" + "zip2 z12.d, z18.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x20, #2, MUL VL]\n" + "zip1 z9.d, z2.d, z13.d\n" + "zip2 z13.d, z2.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x20, #3, MUL VL]\n" + "zip1 z10.d, z17.d, z14.d\n" + "zip2 z14.d, z17.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" "b 15f\n" @@ -290,12 +290,12 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "16:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 17f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" "cbnz x28, 18f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -303,95 +303,95 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "b 18f\n" "17:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" + "add x25, x26, x21\n" "18:" // Height 2: input setup done "cmp x27, #0x10\n" "ble 20f\n" "19:" // Height 2: Multiply loop: Main loop head "whilelt p0.b, XZR, x27\n" - "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + "ld1rqb { z20.b }, p0/Z, [x26]\n" + "ld1rqb { z19.b }, p0/Z, [x25]\n" + "trn1 z18.d, z20.d, z19.d\n" + "ld1b { z17.b }, p5/Z, [x10]\n" + "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d19a48 // ummla z8.s, z18.b, z17.b\n" + ".inst 0x45d09a4c // ummla z12.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d19a49 // ummla z9.s, z18.b, z17.b\n" + ".inst 0x45d09a4d // ummla z13.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45d19a4a // ummla z10.s, z18.b, z17.b\n" + ".inst 0x45d09a4e // ummla z14.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - "trn2 z1.d, z1.d, z2.d\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-8, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #-7, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-6, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #-5, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-4, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #-3, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #-1, MUL VL]\n" + "trn2 z20.d, z20.d, z19.d\n" + ".inst 0x45d19a4b // ummla z11.s, z18.b, z17.b\n" + ".inst 0x45d09a4f // ummla z15.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #-8, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #-7, MUL VL]\n" + ".inst 0x45d19a88 // ummla z8.s, z20.b, z17.b\n" + ".inst 0x45d09a8c // ummla z12.s, z20.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #-6, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #-5, MUL VL]\n" + ".inst 0x45d19a89 // ummla z9.s, z20.b, z17.b\n" + ".inst 0x45d09a8d // ummla z13.s, z20.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #-4, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #-3, MUL VL]\n" + ".inst 0x45d19a8a // ummla z10.s, z20.b, z17.b\n" + ".inst 0x45d09a8e // ummla z14.s, z20.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #-2, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #-1, MUL VL]\n" "sub x27, x27, #0x10\n" "cmp x27, #0x10\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" + ".inst 0x45d19a8b // ummla z11.s, z20.b, z17.b\n" + ".inst 0x45d09a8f // ummla z15.s, z20.b, z16.b\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "bgt 19b\n" "20:" // Height 2: Multiply loop: Single iteration only "whilelt p0.b, XZR, x27\n" "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + "ld1rqb { z19.b }, p0/Z, [x25]\n" + "trn1 z18.d, z1.d, z19.d\n" + "ld1b { z17.b }, p5/Z, [x10]\n" + "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d19a48 // ummla z8.s, z18.b, z17.b\n" + ".inst 0x45d09a4c // ummla z12.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d19a49 // ummla z9.s, z18.b, z17.b\n" + ".inst 0x45d09a4d // ummla z13.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45d19a4a // ummla z10.s, z18.b, z17.b\n" + ".inst 0x45d09a4e // ummla z14.s, z18.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n" "subs x27, x27, #0x8\n" - "trn2 z1.d, z1.d, z2.d\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" + "trn2 z1.d, z1.d, z19.d\n" + ".inst 0x45d19a4b // ummla z11.s, z18.b, z17.b\n" + ".inst 0x45d09a4f // ummla z15.s, z18.b, z16.b\n" "addvl x10, x10, #8\n" "ble 21f\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" + "ld1b { z17.b }, p5/Z, [x10]\n" + "ld1b { z16.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d19828 // ummla z8.s, z1.b, z17.b\n" + ".inst 0x45d0982c // ummla z12.s, z1.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d19829 // ummla z9.s, z1.b, z17.b\n" + ".inst 0x45d0982d // ummla z13.s, z1.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #4, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45d1982a // ummla z10.s, z1.b, z17.b\n" + ".inst 0x45d0982e // ummla z14.s, z1.b, z16.b\n" + "ld1b { z17.b }, p5/Z, [x10, #6, MUL VL]\n" + "ld1b { z16.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45d1982b // ummla z11.s, z1.b, z17.b\n" + ".inst 0x45d0982f // ummla z15.s, z1.b, z16.b\n" "addvl x10, x10, #8\n" "21:" // Height 2: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -399,24 +399,24 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "cmp x28, x20\n" "bne 16b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "uzp1 z7.d, z8.d, z12.d\n" + "add x20, x9, x20, LSL #2\n" + "uzp1 z16.d, z8.d, z12.d\n" "uzp2 z8.d, z8.d, z12.d\n" - "uzp1 z12.d, z9.d, z13.d\n" + "uzp1 z17.d, z9.d, z13.d\n" "uzp2 z9.d, z9.d, z13.d\n" - "st1w { z7.s }, p4, [x9]\n" - "uzp1 z13.d, z10.d, z14.d\n" + "st1w { z16.s }, p4, [x9]\n" + "uzp1 z16.d, z10.d, z14.d\n" "uzp2 z10.d, z10.d, z14.d\n" - "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" - "uzp1 z14.d, z11.d, z15.d\n" + "st1w { z17.s }, p3, [x9, #1, MUL VL]\n" + "uzp1 z2.d, z11.d, z15.d\n" "uzp2 z11.d, z11.d, z15.d\n" - "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" - "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" + "st1w { z16.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z2.s }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" - "st1w { z8.s }, p4, [x24]\n" - "st1w { z9.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z8.s }, p4, [x20]\n" + "st1w { z9.s }, p3, [x20, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x20, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x20, #3, MUL VL]\n" "22:" // Height 2: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -437,28 +437,28 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "whilelt p1.s, x20, x11\n" "tbz %x[flags], #0, 25f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "add x21, x9, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z19.s }, p4/Z, [x9]\n" + "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x24]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x23]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x21]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x21, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z13.d\n" + "zip2 z13.d, z17.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x20]\n" + "zip1 z10.d, z18.d, z14.d\n" + "zip2 z14.d, z18.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z24.s }, p1/Z, [x20, #3, MUL VL]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" "zip1 z17.d, z18.d, z21.d\n" @@ -490,13 +490,13 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "27:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 28f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" "cbnz x28, 29f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -505,169 +505,169 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "b 29f\n" "28:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" "29:" // Height 3: input setup done "cmp x27, #0x10\n" "ble 31f\n" "30:" // Height 3: Multiply loop: Main loop head "whilelt p0.b, XZR, x27\n" - "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "trn1 z0.d, z1.d, z2.d\n" - "trn2 z1.d, z1.d, z2.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c79850 // ummla z16.s, z2.b, z7.b\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - ".inst 0x45c69854 // ummla z20.s, z2.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c79851 // ummla z17.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - "trn2 z3.d, z3.d, z4.d\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - ".inst 0x45c69855 // ummla z21.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" + "ld1rqb { z30.b }, p0/Z, [x26]\n" + "ld1rqb { z24.b }, p0/Z, [x25]\n" + "ld1rqb { z28.b }, p0/Z, [x24]\n" + "trn1 z27.d, z30.d, z24.d\n" + "trn2 z30.d, z30.d, z24.d\n" + "ld1b { z25.b }, p5/Z, [x10]\n" + "trn1 z26.d, z28.d, z29.d\n" + "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d99b68 // ummla z8.s, z27.b, z25.b\n" + ".inst 0x45d99b50 // ummla z16.s, z26.b, z25.b\n" + ".inst 0x45d89b6c // ummla z12.s, z27.b, z24.b\n" + ".inst 0x45d89b54 // ummla z20.s, z26.b, z24.b\n" + "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d99b69 // ummla z9.s, z27.b, z25.b\n" + ".inst 0x45d99b51 // ummla z17.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n" + "trn2 z28.d, z28.d, z29.d\n" + ".inst 0x45d89b6d // ummla z13.s, z27.b, z24.b\n" + ".inst 0x45d89b55 // ummla z21.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n" "sub x27, x27, #0x10\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c79852 // ummla z18.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45d99b6a // ummla z10.s, z27.b, z25.b\n" + ".inst 0x45d99b52 // ummla z18.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n" "cmp x27, #0x10\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - ".inst 0x45c69856 // ummla z22.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45d89b6e // ummla z14.s, z27.b, z24.b\n" + ".inst 0x45d89b56 // ummla z22.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c79853 // ummla z19.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-8, MUL VL]\n" + ".inst 0x45d99b6b // ummla z11.s, z27.b, z25.b\n" + ".inst 0x45d99b53 // ummla z19.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n" "add x26, x26, #0x10\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - ".inst 0x45c69857 // ummla z23.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-7, MUL VL]\n" + ".inst 0x45d89b6f // ummla z15.s, z27.b, z24.b\n" + ".inst 0x45d89b57 // ummla z23.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n" "add x25, x25, #0x10\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c79870 // ummla z16.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-6, MUL VL]\n" + ".inst 0x45d99bc8 // ummla z8.s, z30.b, z25.b\n" + ".inst 0x45d99b90 // ummla z16.s, z28.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - ".inst 0x45c69874 // ummla z20.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-5, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c79871 // ummla z17.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-4, MUL VL]\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - ".inst 0x45c69875 // ummla z21.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-3, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c79872 // ummla z18.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-2, MUL VL]\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - ".inst 0x45c69876 // ummla z22.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-1, MUL VL]\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c79873 // ummla z19.s, z3.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" - ".inst 0x45c69877 // ummla z23.s, z3.b, z6.b\n" + ".inst 0x45d89bcc // ummla z12.s, z30.b, z24.b\n" + ".inst 0x45d89b94 // ummla z20.s, z28.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n" + ".inst 0x45d99bc9 // ummla z9.s, z30.b, z25.b\n" + ".inst 0x45d99b91 // ummla z17.s, z28.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n" + ".inst 0x45d89bcd // ummla z13.s, z30.b, z24.b\n" + ".inst 0x45d89b95 // ummla z21.s, z28.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n" + ".inst 0x45d99bca // ummla z10.s, z30.b, z25.b\n" + ".inst 0x45d99b92 // ummla z18.s, z28.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n" + ".inst 0x45d89bce // ummla z14.s, z30.b, z24.b\n" + ".inst 0x45d89b96 // ummla z22.s, z28.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n" + ".inst 0x45d99bcb // ummla z11.s, z30.b, z25.b\n" + ".inst 0x45d99b93 // ummla z19.s, z28.b, z25.b\n" + ".inst 0x45d89bcf // ummla z15.s, z30.b, z24.b\n" + ".inst 0x45d89b97 // ummla z23.s, z28.b, z24.b\n" "bgt 30b\n" "31:" // Height 3: Multiply loop: Single iteration only "whilelt p0.b, XZR, x27\n" "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" + "ld1rqb { z24.b }, p0/Z, [x25]\n" "ld1rqb { z3.b }, p0/Z, [x24]\n" - "trn1 z0.d, z1.d, z2.d\n" - "trn2 z1.d, z1.d, z2.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c79850 // ummla z16.s, z2.b, z7.b\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - ".inst 0x45c69854 // ummla z20.s, z2.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c79851 // ummla z17.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" + "trn1 z27.d, z1.d, z24.d\n" + "trn2 z1.d, z1.d, z24.d\n" + "ld1b { z25.b }, p5/Z, [x10]\n" + "trn1 z26.d, z3.d, z28.d\n" + "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d99b68 // ummla z8.s, z27.b, z25.b\n" + ".inst 0x45d99b50 // ummla z16.s, z26.b, z25.b\n" + ".inst 0x45d89b6c // ummla z12.s, z27.b, z24.b\n" + ".inst 0x45d89b54 // ummla z20.s, z26.b, z24.b\n" + "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d99b69 // ummla z9.s, z27.b, z25.b\n" + ".inst 0x45d99b51 // ummla z17.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n" "subs x27, x27, #0x8\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - ".inst 0x45c69855 // ummla z21.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - "trn2 z3.d, z3.d, z4.d\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c79852 // ummla z18.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - ".inst 0x45c69856 // ummla z22.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45d89b6d // ummla z13.s, z27.b, z24.b\n" + ".inst 0x45d89b55 // ummla z21.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n" + "trn2 z3.d, z3.d, z28.d\n" + ".inst 0x45d99b6a // ummla z10.s, z27.b, z25.b\n" + ".inst 0x45d99b52 // ummla z18.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45d89b6e // ummla z14.s, z27.b, z24.b\n" + ".inst 0x45d89b56 // ummla z22.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c79853 // ummla z19.s, z2.b, z7.b\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - ".inst 0x45c69857 // ummla z23.s, z2.b, z6.b\n" + ".inst 0x45d99b6b // ummla z11.s, z27.b, z25.b\n" + ".inst 0x45d99b53 // ummla z19.s, z26.b, z25.b\n" + ".inst 0x45d89b6f // ummla z15.s, z27.b, z24.b\n" + ".inst 0x45d89b57 // ummla z23.s, z26.b, z24.b\n" "ble 32f\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c79870 // ummla z16.s, z3.b, z7.b\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - ".inst 0x45c69874 // ummla z20.s, z3.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c79871 // ummla z17.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - ".inst 0x45c69875 // ummla z21.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c79872 // ummla z18.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - ".inst 0x45c69876 // ummla z22.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + "ld1b { z25.b }, p5/Z, [x10]\n" + "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d99828 // ummla z8.s, z1.b, z25.b\n" + ".inst 0x45d99870 // ummla z16.s, z3.b, z25.b\n" + ".inst 0x45d8982c // ummla z12.s, z1.b, z24.b\n" + ".inst 0x45d89874 // ummla z20.s, z3.b, z24.b\n" + "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d99829 // ummla z9.s, z1.b, z25.b\n" + ".inst 0x45d99871 // ummla z17.s, z3.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x45d8982d // ummla z13.s, z1.b, z24.b\n" + ".inst 0x45d89875 // ummla z21.s, z3.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45d9982a // ummla z10.s, z1.b, z25.b\n" + ".inst 0x45d99872 // ummla z18.s, z3.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45d8982e // ummla z14.s, z1.b, z24.b\n" + ".inst 0x45d89876 // ummla z22.s, z3.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c79873 // ummla z19.s, z3.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" - ".inst 0x45c69877 // ummla z23.s, z3.b, z6.b\n" + ".inst 0x45d9982b // ummla z11.s, z1.b, z25.b\n" + ".inst 0x45d99873 // ummla z19.s, z3.b, z25.b\n" + ".inst 0x45d8982f // ummla z15.s, z1.b, z24.b\n" + ".inst 0x45d89877 // ummla z23.s, z3.b, z24.b\n" "32:" // Height 3: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" "bne 27b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "uzp1 z7.d, z8.d, z12.d\n" + "add x21, x9, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" + "uzp1 z25.d, z8.d, z12.d\n" "uzp2 z8.d, z8.d, z12.d\n" - "uzp1 z12.d, z9.d, z13.d\n" - "st1w { z7.s }, p4, [x9]\n" + "uzp1 z24.d, z9.d, z13.d\n" + "st1w { z25.s }, p4, [x9]\n" "uzp2 z9.d, z9.d, z13.d\n" - "uzp1 z13.d, z10.d, z14.d\n" - "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" + "uzp1 z25.d, z10.d, z14.d\n" + "st1w { z24.s }, p3, [x9, #1, MUL VL]\n" "uzp2 z10.d, z10.d, z14.d\n" - "uzp1 z14.d, z11.d, z15.d\n" - "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" + "uzp1 z24.d, z11.d, z15.d\n" + "st1w { z25.s }, p2, [x9, #2, MUL VL]\n" "uzp2 z11.d, z11.d, z15.d\n" "uzp1 z16.d, z16.d, z20.d\n" - "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" + "st1w { z24.s }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" "uzp1 z17.d, z17.d, z21.d\n" "uzp1 z18.d, z18.d, z22.d\n" - "st1w { z8.s }, p4, [x24]\n" + "st1w { z8.s }, p4, [x21]\n" "uzp1 z19.d, z19.d, z23.d\n" - "st1w { z9.s }, p3, [x24, #1, MUL VL]\n" - "st1w { z10.s }, p2, [x24, #2, MUL VL]\n" - "st1w { z11.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x23]\n" - "st1w { z17.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x23, #3, MUL VL]\n" + "st1w { z9.s }, p3, [x21, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x21, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x21, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x20]\n" + "st1w { z17.s }, p3, [x20, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x20, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x20, #3, MUL VL]\n" "33:" // Height 3: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -688,37 +688,37 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "whilelt p1.s, x20, x11\n" "tbz %x[flags], #0, 36f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" - "add x22, x23, x20, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "add x22, x9, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" + "ld1w { z19.s }, p4/Z, [x9]\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x24]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x23]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x22]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x22, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z13.d\n" + "zip2 z13.d, z17.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x21]\n" + "zip1 z10.d, z18.d, z14.d\n" + "zip2 z14.d, z18.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x21, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x22]\n" + "ld1w { z24.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x20]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" - "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z21.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z17.d, z18.d, z21.d\n" "zip2 z21.d, z18.d, z21.d\n" - "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x20, #3, MUL VL]\n" "zip1 z18.d, z19.d, z22.d\n" "zip2 z22.d, z19.d, z22.d\n" "zip1 z19.d, z24.d, z23.d\n" @@ -746,14 +746,14 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "38:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 39f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" "cbnz x28, 40f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -763,182 +763,182 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "b 40f\n" "39:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" + "add x23, x24, x21\n" "40:" // Height 4: input setup done "cmp x27, #0x10\n" "ble 42f\n" "41:" // Height 4: Multiply loop: Main loop head "whilelt p0.b, XZR, x27\n" - "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "ld1rqb { z4.b }, p0/Z, [x23]\n" - "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c79850 // ummla z16.s, z2.b, z7.b\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - ".inst 0x45c69854 // ummla z20.s, z2.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c79851 // ummla z17.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - "trn2 z3.d, z3.d, z4.d\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - ".inst 0x45c69855 // ummla z21.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" + "ld1rqb { z30.b }, p0/Z, [x26]\n" + "ld1rqb { z24.b }, p0/Z, [x25]\n" + "trn1 z29.d, z30.d, z24.d\n" + "ld1rqb { z28.b }, p0/Z, [x24]\n" + "ld1rqb { z27.b }, p0/Z, [x23]\n" + "trn2 z30.d, z30.d, z24.d\n" + "trn1 z26.d, z28.d, z27.d\n" + "ld1b { z25.b }, p5/Z, [x10]\n" + "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d99ba8 // ummla z8.s, z29.b, z25.b\n" + ".inst 0x45d99b50 // ummla z16.s, z26.b, z25.b\n" + ".inst 0x45d89bac // ummla z12.s, z29.b, z24.b\n" + ".inst 0x45d89b54 // ummla z20.s, z26.b, z24.b\n" + "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d99ba9 // ummla z9.s, z29.b, z25.b\n" + ".inst 0x45d99b51 // ummla z17.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n" + "trn2 z28.d, z28.d, z27.d\n" + ".inst 0x45d89bad // ummla z13.s, z29.b, z24.b\n" + ".inst 0x45d89b55 // ummla z21.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n" "sub x27, x27, #0x10\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c79852 // ummla z18.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45d99baa // ummla z10.s, z29.b, z25.b\n" + ".inst 0x45d99b52 // ummla z18.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n" "cmp x27, #0x10\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - ".inst 0x45c69856 // ummla z22.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45d89bae // ummla z14.s, z29.b, z24.b\n" + ".inst 0x45d89b56 // ummla z22.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c79853 // ummla z19.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-8, MUL VL]\n" + ".inst 0x45d99bab // ummla z11.s, z29.b, z25.b\n" + ".inst 0x45d99b53 // ummla z19.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #-8, MUL VL]\n" "add x26, x26, #0x10\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - ".inst 0x45c69857 // ummla z23.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-7, MUL VL]\n" + ".inst 0x45d89baf // ummla z15.s, z29.b, z24.b\n" + ".inst 0x45d89b57 // ummla z23.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #-7, MUL VL]\n" "add x25, x25, #0x10\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c79870 // ummla z16.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-6, MUL VL]\n" + ".inst 0x45d99bc8 // ummla z8.s, z30.b, z25.b\n" + ".inst 0x45d99b90 // ummla z16.s, z28.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #-6, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - ".inst 0x45c69874 // ummla z20.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-5, MUL VL]\n" + ".inst 0x45d89bcc // ummla z12.s, z30.b, z24.b\n" + ".inst 0x45d89b94 // ummla z20.s, z28.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #-5, MUL VL]\n" "add x23, x23, #0x10\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c79871 // ummla z17.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-4, MUL VL]\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - ".inst 0x45c69875 // ummla z21.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-3, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c79872 // ummla z18.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-2, MUL VL]\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - ".inst 0x45c69876 // ummla z22.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-1, MUL VL]\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c79873 // ummla z19.s, z3.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" - ".inst 0x45c69877 // ummla z23.s, z3.b, z6.b\n" + ".inst 0x45d99bc9 // ummla z9.s, z30.b, z25.b\n" + ".inst 0x45d99b91 // ummla z17.s, z28.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #-4, MUL VL]\n" + ".inst 0x45d89bcd // ummla z13.s, z30.b, z24.b\n" + ".inst 0x45d89b95 // ummla z21.s, z28.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #-3, MUL VL]\n" + ".inst 0x45d99bca // ummla z10.s, z30.b, z25.b\n" + ".inst 0x45d99b92 // ummla z18.s, z28.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #-2, MUL VL]\n" + ".inst 0x45d89bce // ummla z14.s, z30.b, z24.b\n" + ".inst 0x45d89b96 // ummla z22.s, z28.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #-1, MUL VL]\n" + ".inst 0x45d99bcb // ummla z11.s, z30.b, z25.b\n" + ".inst 0x45d99b93 // ummla z19.s, z28.b, z25.b\n" + ".inst 0x45d89bcf // ummla z15.s, z30.b, z24.b\n" + ".inst 0x45d89b97 // ummla z23.s, z28.b, z24.b\n" "bgt 41b\n" "42:" // Height 4: Multiply loop: Single iteration only "whilelt p0.b, XZR, x27\n" "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" + "ld1rqb { z24.b }, p0/Z, [x25]\n" + "trn1 z28.d, z1.d, z24.d\n" "ld1rqb { z3.b }, p0/Z, [x24]\n" - "ld1rqb { z4.b }, p0/Z, [x23]\n" - "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c79850 // ummla z16.s, z2.b, z7.b\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - ".inst 0x45c69854 // ummla z20.s, z2.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c79851 // ummla z17.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" + "ld1rqb { z27.b }, p0/Z, [x23]\n" + "trn2 z1.d, z1.d, z24.d\n" + "trn1 z26.d, z3.d, z27.d\n" + "ld1b { z25.b }, p5/Z, [x10]\n" + "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d99b88 // ummla z8.s, z28.b, z25.b\n" + ".inst 0x45d99b50 // ummla z16.s, z26.b, z25.b\n" + ".inst 0x45d89b8c // ummla z12.s, z28.b, z24.b\n" + ".inst 0x45d89b54 // ummla z20.s, z26.b, z24.b\n" + "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d99b89 // ummla z9.s, z28.b, z25.b\n" + ".inst 0x45d99b51 // ummla z17.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n" "subs x27, x27, #0x8\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - ".inst 0x45c69855 // ummla z21.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - "trn2 z3.d, z3.d, z4.d\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c79852 // ummla z18.s, z2.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - ".inst 0x45c69856 // ummla z22.s, z2.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45d89b8d // ummla z13.s, z28.b, z24.b\n" + ".inst 0x45d89b55 // ummla z21.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n" + "trn2 z3.d, z3.d, z27.d\n" + ".inst 0x45d99b8a // ummla z10.s, z28.b, z25.b\n" + ".inst 0x45d99b52 // ummla z18.s, z26.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45d89b8e // ummla z14.s, z28.b, z24.b\n" + ".inst 0x45d89b56 // ummla z22.s, z26.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c79853 // ummla z19.s, z2.b, z7.b\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - ".inst 0x45c69857 // ummla z23.s, z2.b, z6.b\n" + ".inst 0x45d99b8b // ummla z11.s, z28.b, z25.b\n" + ".inst 0x45d99b53 // ummla z19.s, z26.b, z25.b\n" + ".inst 0x45d89b8f // ummla z15.s, z28.b, z24.b\n" + ".inst 0x45d89b57 // ummla z23.s, z26.b, z24.b\n" "ble 43f\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c79870 // ummla z16.s, z3.b, z7.b\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - ".inst 0x45c69874 // ummla z20.s, z3.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c79871 // ummla z17.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - ".inst 0x45c69875 // ummla z21.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c79872 // ummla z18.s, z3.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - ".inst 0x45c69876 // ummla z22.s, z3.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + "ld1b { z25.b }, p5/Z, [x10]\n" + "ld1b { z24.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45d99828 // ummla z8.s, z1.b, z25.b\n" + ".inst 0x45d99870 // ummla z16.s, z3.b, z25.b\n" + ".inst 0x45d8982c // ummla z12.s, z1.b, z24.b\n" + ".inst 0x45d89874 // ummla z20.s, z3.b, z24.b\n" + "ld1b { z25.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1b { z24.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45d99829 // ummla z9.s, z1.b, z25.b\n" + ".inst 0x45d99871 // ummla z17.s, z3.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x45d8982d // ummla z13.s, z1.b, z24.b\n" + ".inst 0x45d89875 // ummla z21.s, z3.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45d9982a // ummla z10.s, z1.b, z25.b\n" + ".inst 0x45d99872 // ummla z18.s, z3.b, z25.b\n" + "ld1b { z25.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45d8982e // ummla z14.s, z1.b, z24.b\n" + ".inst 0x45d89876 // ummla z22.s, z3.b, z24.b\n" + "ld1b { z24.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c79873 // ummla z19.s, z3.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" - ".inst 0x45c69877 // ummla z23.s, z3.b, z6.b\n" + ".inst 0x45d9982b // ummla z11.s, z1.b, z25.b\n" + ".inst 0x45d99873 // ummla z19.s, z3.b, z25.b\n" + ".inst 0x45d8982f // ummla z15.s, z1.b, z24.b\n" + ".inst 0x45d89877 // ummla z23.s, z3.b, z24.b\n" "43:" // Height 4: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" "bne 38b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "uzp1 z7.d, z8.d, z12.d\n" - "add x22, x23, x20, LSL #2\n" + "add x22, x9, x20, LSL #2\n" + "add x21, x22, x20, LSL #2\n" + "uzp1 z25.d, z8.d, z12.d\n" + "add x20, x21, x20, LSL #2\n" "uzp2 z8.d, z8.d, z12.d\n" - "uzp1 z12.d, z9.d, z13.d\n" - "st1w { z7.s }, p4, [x9]\n" + "uzp1 z24.d, z9.d, z13.d\n" + "st1w { z25.s }, p4, [x9]\n" "uzp2 z9.d, z9.d, z13.d\n" - "uzp1 z13.d, z10.d, z14.d\n" - "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" + "uzp1 z25.d, z10.d, z14.d\n" + "st1w { z24.s }, p3, [x9, #1, MUL VL]\n" "uzp2 z10.d, z10.d, z14.d\n" - "uzp1 z14.d, z11.d, z15.d\n" - "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" + "uzp1 z24.d, z11.d, z15.d\n" + "st1w { z25.s }, p2, [x9, #2, MUL VL]\n" "uzp2 z11.d, z11.d, z15.d\n" - "uzp1 z15.d, z16.d, z20.d\n" - "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" + "uzp1 z25.d, z16.d, z20.d\n" + "st1w { z24.s }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" "uzp2 z16.d, z16.d, z20.d\n" - "uzp1 z20.d, z17.d, z21.d\n" - "st1w { z8.s }, p4, [x24]\n" + "uzp1 z24.d, z17.d, z21.d\n" + "st1w { z8.s }, p4, [x22]\n" "uzp2 z17.d, z17.d, z21.d\n" "uzp1 z21.d, z18.d, z22.d\n" - "st1w { z9.s }, p3, [x24, #1, MUL VL]\n" + "st1w { z9.s }, p3, [x22, #1, MUL VL]\n" "uzp2 z18.d, z18.d, z22.d\n" - "uzp1 z22.d, z19.d, z23.d\n" - "st1w { z10.s }, p2, [x24, #2, MUL VL]\n" + "uzp1 z20.d, z19.d, z23.d\n" + "st1w { z10.s }, p2, [x22, #2, MUL VL]\n" "uzp2 z19.d, z19.d, z23.d\n" - "st1w { z11.s }, p1, [x24, #3, MUL VL]\n" - "st1w { z15.s }, p4, [x23]\n" - "st1w { z20.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z21.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z22.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x22]\n" - "st1w { z17.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x22, #3, MUL VL]\n" + "st1w { z11.s }, p1, [x22, #3, MUL VL]\n" + "st1w { z25.s }, p4, [x21]\n" + "st1w { z24.s }, p3, [x21, #1, MUL VL]\n" + "st1w { z21.s }, p2, [x21, #2, MUL VL]\n" + "st1w { z20.s }, p1, [x21, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x20]\n" + "st1w { z17.s }, p3, [x20, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x20, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x20, #3, MUL VL]\n" "44:" // Height 4: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -959,54 +959,54 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "whilelt p1.s, x20, x11\n" "tbz %x[flags], #0, 47f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" + "add x23, x9, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" + "ld1w { z19.s }, p4/Z, [x9]\n" "add x21, x22, x20, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "add x20, x21, x20, LSL #2\n" + "ld1w { z17.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x9, #2, MUL VL]\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" - "ld1w { z12.s }, p4/Z, [x24]\n" - "zip1 z8.d, z9.d, z12.d\n" - "zip2 z12.d, z9.d, z12.d\n" - "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" - "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "zip1 z9.d, z10.d, z13.d\n" - "zip2 z13.d, z10.d, z13.d\n" - "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" - "ld1w { z17.s }, p4/Z, [x23]\n" - "zip1 z10.d, z11.d, z14.d\n" - "zip2 z14.d, z11.d, z14.d\n" - "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" - "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x23]\n" + "zip1 z8.d, z19.d, z12.d\n" + "zip2 z12.d, z19.d, z12.d\n" + "ld1w { z13.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x23, #2, MUL VL]\n" + "zip1 z9.d, z17.d, z13.d\n" + "zip2 z13.d, z17.d, z13.d\n" + "ld1w { z15.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z17.s }, p4/Z, [x22]\n" + "zip1 z10.d, z18.d, z14.d\n" + "zip2 z14.d, z18.d, z14.d\n" + "ld1w { z18.s }, p3/Z, [x22, #1, MUL VL]\n" + "ld1w { z19.s }, p2/Z, [x22, #2, MUL VL]\n" "zip1 z11.d, z16.d, z15.d\n" "zip2 z15.d, z16.d, z15.d\n" - "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" - "ld1w { z20.s }, p4/Z, [x22]\n" + "ld1w { z24.s }, p1/Z, [x22, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x21]\n" "zip1 z16.d, z17.d, z20.d\n" "zip2 z20.d, z17.d, z20.d\n" - "ld1w { z21.s }, p3/Z, [x22, #1, MUL VL]\n" - "ld1w { z22.s }, p2/Z, [x22, #2, MUL VL]\n" + "ld1w { z21.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x21, #2, MUL VL]\n" "zip1 z17.d, z18.d, z21.d\n" "zip2 z21.d, z18.d, z21.d\n" - "ld1w { z23.s }, p1/Z, [x22, #3, MUL VL]\n" - "ld1w { z25.s }, p4/Z, [x21]\n" + "ld1w { z23.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z25.s }, p4/Z, [x20]\n" "zip1 z18.d, z19.d, z22.d\n" "zip2 z22.d, z19.d, z22.d\n" - "ld1w { z26.s }, p3/Z, [x21, #1, MUL VL]\n" - "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z26.s }, p3/Z, [x20, #1, MUL VL]\n" + "ld1w { z27.s }, p2/Z, [x20, #2, MUL VL]\n" "zip1 z19.d, z24.d, z23.d\n" "zip2 z23.d, z24.d, z23.d\n" - "ld1w { z6.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z0.s }, p1/Z, [x20, #3, MUL VL]\n" "zip1 z24.d, z25.d, z28.d\n" "zip2 z28.d, z25.d, z28.d\n" "zip1 z25.d, z26.d, z29.d\n" "zip2 z29.d, z26.d, z29.d\n" "zip1 z26.d, z27.d, z30.d\n" "zip2 z30.d, z27.d, z30.d\n" - "zip1 z27.d, z6.d, z31.d\n" - "zip2 z31.d, z6.d, z31.d\n" + "zip1 z27.d, z0.d, z31.d\n" + "zip2 z31.d, z0.d, z31.d\n" "b 48f\n" "47:" // Height 5: no accumulate "mov z8.s, #0x0\n" @@ -1038,15 +1038,15 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "49:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 50f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" "cbnz x28, 51f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -1057,231 +1057,231 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "b 51f\n" "50:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" - "add x22, x23, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" + "add x23, x24, x21\n" + "add x22, x23, x21\n" "51:" // Height 5: input setup done "cmp x27, #0x10\n" "ble 53f\n" "52:" // Height 5: Multiply loop: Main loop head "whilelt p0.b, XZR, x27\n" - "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "ld1rqb { z4.b }, p0/Z, [x23]\n" - "trn1 z0.d, z1.d, z2.d\n" - "trn2 z1.d, z1.d, z2.d\n" - "ld1rqb { z5.b }, p0/Z, [x22]\n" - "trn1 z2.d, z3.d, z4.d\n" - "trn2 z3.d, z3.d, z4.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "trn1 z4.d, z5.d, z6.d\n" - "trn2 z5.d, z5.d, z6.d\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c79850 // ummla z16.s, z2.b, z7.b\n" - ".inst 0x45c79898 // ummla z24.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1rqb { z6.b }, p0/Z, [x26]\n" + "ld1rqb { z1.b }, p0/Z, [x25]\n" + "ld1rqb { z7.b }, p0/Z, [x24]\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "trn1 z5.d, z6.d, z1.d\n" + "trn2 z6.d, z6.d, z1.d\n" + "ld1rqb { z4.b }, p0/Z, [x22]\n" + "trn1 z3.d, z7.d, z2.d\n" + "trn2 z7.d, z7.d, z2.d\n" + "ld1b { z1.b }, p5/Z, [x10]\n" + "trn1 z2.d, z4.d, z0.d\n" + "trn2 z4.d, z4.d, z0.d\n" + "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45c198a8 // ummla z8.s, z5.b, z1.b\n" + ".inst 0x45c19870 // ummla z16.s, z3.b, z1.b\n" + ".inst 0x45c19858 // ummla z24.s, z2.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n" "sub x27, x27, #0x10\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - ".inst 0x45c69854 // ummla z20.s, z2.b, z6.b\n" + ".inst 0x45c098ac // ummla z12.s, z5.b, z0.b\n" + ".inst 0x45c09874 // ummla z20.s, z3.b, z0.b\n" "cmp x27, #0x10\n" "add x26, x26, #0x10\n" - ".inst 0x45c6989c // ummla z28.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" + ".inst 0x45c0985c // ummla z28.s, z2.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45c198a9 // ummla z9.s, z5.b, z1.b\n" "add x25, x25, #0x10\n" - ".inst 0x45c79851 // ummla z17.s, z2.b, z7.b\n" - ".inst 0x45c79899 // ummla z25.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x45c19871 // ummla z17.s, z3.b, z1.b\n" + ".inst 0x45c19859 // ummla z25.s, z2.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - ".inst 0x45c69855 // ummla z21.s, z2.b, z6.b\n" + ".inst 0x45c098ad // ummla z13.s, z5.b, z0.b\n" + ".inst 0x45c09875 // ummla z21.s, z3.b, z0.b\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - ".inst 0x45c6989d // ummla z29.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c79852 // ummla z18.s, z2.b, z7.b\n" - ".inst 0x45c7989a // ummla z26.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - ".inst 0x45c69856 // ummla z22.s, z2.b, z6.b\n" - ".inst 0x45c6989e // ummla z30.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45c0985d // ummla z29.s, z2.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45c198aa // ummla z10.s, z5.b, z1.b\n" + ".inst 0x45c19872 // ummla z18.s, z3.b, z1.b\n" + ".inst 0x45c1985a // ummla z26.s, z2.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45c098ae // ummla z14.s, z5.b, z0.b\n" + ".inst 0x45c09876 // ummla z22.s, z3.b, z0.b\n" + ".inst 0x45c0985e // ummla z30.s, z2.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c79853 // ummla z19.s, z2.b, z7.b\n" - ".inst 0x45c7989b // ummla z27.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-8, MUL VL]\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - ".inst 0x45c69857 // ummla z23.s, z2.b, z6.b\n" - ".inst 0x45c6989f // ummla z31.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-7, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c79870 // ummla z16.s, z3.b, z7.b\n" - ".inst 0x45c798b8 // ummla z24.s, z5.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-6, MUL VL]\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - ".inst 0x45c69874 // ummla z20.s, z3.b, z6.b\n" - ".inst 0x45c698bc // ummla z28.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-5, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c79871 // ummla z17.s, z3.b, z7.b\n" - ".inst 0x45c798b9 // ummla z25.s, z5.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-4, MUL VL]\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - ".inst 0x45c69875 // ummla z21.s, z3.b, z6.b\n" - ".inst 0x45c698bd // ummla z29.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-3, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c79872 // ummla z18.s, z3.b, z7.b\n" - ".inst 0x45c798ba // ummla z26.s, z5.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-2, MUL VL]\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - ".inst 0x45c69876 // ummla z22.s, z3.b, z6.b\n" - ".inst 0x45c698be // ummla z30.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-1, MUL VL]\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c79873 // ummla z19.s, z3.b, z7.b\n" - ".inst 0x45c798bb // ummla z27.s, z5.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" - ".inst 0x45c69877 // ummla z23.s, z3.b, z6.b\n" - ".inst 0x45c698bf // ummla z31.s, z5.b, z6.b\n" + ".inst 0x45c198ab // ummla z11.s, z5.b, z1.b\n" + ".inst 0x45c19873 // ummla z19.s, z3.b, z1.b\n" + ".inst 0x45c1985b // ummla z27.s, z2.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n" + ".inst 0x45c098af // ummla z15.s, z5.b, z0.b\n" + ".inst 0x45c09877 // ummla z23.s, z3.b, z0.b\n" + ".inst 0x45c0985f // ummla z31.s, z2.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #-7, MUL VL]\n" + ".inst 0x45c198c8 // ummla z8.s, z6.b, z1.b\n" + ".inst 0x45c198f0 // ummla z16.s, z7.b, z1.b\n" + ".inst 0x45c19898 // ummla z24.s, z4.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #-6, MUL VL]\n" + ".inst 0x45c098cc // ummla z12.s, z6.b, z0.b\n" + ".inst 0x45c098f4 // ummla z20.s, z7.b, z0.b\n" + ".inst 0x45c0989c // ummla z28.s, z4.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #-5, MUL VL]\n" + ".inst 0x45c198c9 // ummla z9.s, z6.b, z1.b\n" + ".inst 0x45c198f1 // ummla z17.s, z7.b, z1.b\n" + ".inst 0x45c19899 // ummla z25.s, z4.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #-4, MUL VL]\n" + ".inst 0x45c098cd // ummla z13.s, z6.b, z0.b\n" + ".inst 0x45c098f5 // ummla z21.s, z7.b, z0.b\n" + ".inst 0x45c0989d // ummla z29.s, z4.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #-3, MUL VL]\n" + ".inst 0x45c198ca // ummla z10.s, z6.b, z1.b\n" + ".inst 0x45c198f2 // ummla z18.s, z7.b, z1.b\n" + ".inst 0x45c1989a // ummla z26.s, z4.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #-2, MUL VL]\n" + ".inst 0x45c098ce // ummla z14.s, z6.b, z0.b\n" + ".inst 0x45c098f6 // ummla z22.s, z7.b, z0.b\n" + ".inst 0x45c0989e // ummla z30.s, z4.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #-1, MUL VL]\n" + ".inst 0x45c198cb // ummla z11.s, z6.b, z1.b\n" + ".inst 0x45c198f3 // ummla z19.s, z7.b, z1.b\n" + ".inst 0x45c1989b // ummla z27.s, z4.b, z1.b\n" + ".inst 0x45c098cf // ummla z15.s, z6.b, z0.b\n" + ".inst 0x45c098f7 // ummla z23.s, z7.b, z0.b\n" + ".inst 0x45c0989f // ummla z31.s, z4.b, z0.b\n" "bgt 52b\n" "53:" // Height 5: Multiply loop: Single iteration only "whilelt p0.b, XZR, x27\n" "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" + "ld1rqb { z4.b }, p0/Z, [x25]\n" "ld1rqb { z3.b }, p0/Z, [x24]\n" - "ld1rqb { z4.b }, p0/Z, [x23]\n" - "trn1 z0.d, z1.d, z2.d\n" - "trn2 z1.d, z1.d, z2.d\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "trn1 z7.d, z1.d, z4.d\n" + "trn2 z1.d, z1.d, z4.d\n" "ld1rqb { z5.b }, p0/Z, [x22]\n" - "trn1 z2.d, z3.d, z4.d\n" - "trn2 z3.d, z3.d, z4.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "trn1 z4.d, z5.d, z6.d\n" - "trn2 z5.d, z5.d, z6.d\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c79850 // ummla z16.s, z2.b, z7.b\n" - ".inst 0x45c79898 // ummla z24.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" + "trn1 z6.d, z3.d, z2.d\n" + "trn2 z3.d, z3.d, z2.d\n" + "ld1b { z2.b }, p5/Z, [x10]\n" + "trn1 z4.d, z5.d, z0.d\n" + "trn2 z5.d, z5.d, z0.d\n" + "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45c298e8 // ummla z8.s, z7.b, z2.b\n" + ".inst 0x45c298d0 // ummla z16.s, z6.b, z2.b\n" + ".inst 0x45c29898 // ummla z24.s, z4.b, z2.b\n" + "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n" "subs x27, x27, #0x8\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - ".inst 0x45c69854 // ummla z20.s, z2.b, z6.b\n" - ".inst 0x45c6989c // ummla z28.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c79851 // ummla z17.s, z2.b, z7.b\n" - ".inst 0x45c79899 // ummla z25.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - ".inst 0x45c69855 // ummla z21.s, z2.b, z6.b\n" - ".inst 0x45c6989d // ummla z29.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c79852 // ummla z18.s, z2.b, z7.b\n" - ".inst 0x45c7989a // ummla z26.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - ".inst 0x45c69856 // ummla z22.s, z2.b, z6.b\n" - ".inst 0x45c6989e // ummla z30.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" + ".inst 0x45c098ec // ummla z12.s, z7.b, z0.b\n" + ".inst 0x45c098d4 // ummla z20.s, z6.b, z0.b\n" + ".inst 0x45c0989c // ummla z28.s, z4.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45c298e9 // ummla z9.s, z7.b, z2.b\n" + ".inst 0x45c298d1 // ummla z17.s, z6.b, z2.b\n" + ".inst 0x45c29899 // ummla z25.s, z4.b, z2.b\n" + "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x45c098ed // ummla z13.s, z7.b, z0.b\n" + ".inst 0x45c098d5 // ummla z21.s, z6.b, z0.b\n" + ".inst 0x45c0989d // ummla z29.s, z4.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45c298ea // ummla z10.s, z7.b, z2.b\n" + ".inst 0x45c298d2 // ummla z18.s, z6.b, z2.b\n" + ".inst 0x45c2989a // ummla z26.s, z4.b, z2.b\n" + "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45c098ee // ummla z14.s, z7.b, z0.b\n" + ".inst 0x45c098d6 // ummla z22.s, z6.b, z0.b\n" + ".inst 0x45c0989e // ummla z30.s, z4.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45c298eb // ummla z11.s, z7.b, z2.b\n" "addvl x10, x10, #8\n" - ".inst 0x45c79853 // ummla z19.s, z2.b, z7.b\n" - ".inst 0x45c7989b // ummla z27.s, z4.b, z7.b\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - ".inst 0x45c69857 // ummla z23.s, z2.b, z6.b\n" - ".inst 0x45c6989f // ummla z31.s, z4.b, z6.b\n" + ".inst 0x45c298d3 // ummla z19.s, z6.b, z2.b\n" + ".inst 0x45c2989b // ummla z27.s, z4.b, z2.b\n" + ".inst 0x45c098ef // ummla z15.s, z7.b, z0.b\n" + ".inst 0x45c098d7 // ummla z23.s, z6.b, z0.b\n" + ".inst 0x45c0989f // ummla z31.s, z4.b, z0.b\n" "ble 54f\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c79870 // ummla z16.s, z3.b, z7.b\n" - ".inst 0x45c798b8 // ummla z24.s, z5.b, z7.b\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x45c69874 // ummla z20.s, z3.b, z6.b\n" - ".inst 0x45c698bc // ummla z28.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c79871 // ummla z17.s, z3.b, z7.b\n" - ".inst 0x45c798b9 // ummla z25.s, z5.b, z7.b\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x45c69875 // ummla z21.s, z3.b, z6.b\n" - ".inst 0x45c698bd // ummla z29.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c79872 // ummla z18.s, z3.b, z7.b\n" - ".inst 0x45c798ba // ummla z26.s, z5.b, z7.b\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c69876 // ummla z22.s, z3.b, z6.b\n" - ".inst 0x45c698be // ummla z30.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + "ld1b { z2.b }, p5/Z, [x10]\n" + "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45c29828 // ummla z8.s, z1.b, z2.b\n" + ".inst 0x45c29870 // ummla z16.s, z3.b, z2.b\n" + ".inst 0x45c298b8 // ummla z24.s, z5.b, z2.b\n" + ".inst 0x45c0982c // ummla z12.s, z1.b, z0.b\n" + "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x45c09874 // ummla z20.s, z3.b, z0.b\n" + ".inst 0x45c098bc // ummla z28.s, z5.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45c29829 // ummla z9.s, z1.b, z2.b\n" + ".inst 0x45c29871 // ummla z17.s, z3.b, z2.b\n" + ".inst 0x45c298b9 // ummla z25.s, z5.b, z2.b\n" + ".inst 0x45c0982d // ummla z13.s, z1.b, z0.b\n" + "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x45c09875 // ummla z21.s, z3.b, z0.b\n" + ".inst 0x45c098bd // ummla z29.s, z5.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45c2982a // ummla z10.s, z1.b, z2.b\n" + ".inst 0x45c29872 // ummla z18.s, z3.b, z2.b\n" + ".inst 0x45c298ba // ummla z26.s, z5.b, z2.b\n" + ".inst 0x45c0982e // ummla z14.s, z1.b, z0.b\n" + "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45c09876 // ummla z22.s, z3.b, z0.b\n" + ".inst 0x45c098be // ummla z30.s, z5.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c79873 // ummla z19.s, z3.b, z7.b\n" - ".inst 0x45c798bb // ummla z27.s, z5.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" - ".inst 0x45c69877 // ummla z23.s, z3.b, z6.b\n" - ".inst 0x45c698bf // ummla z31.s, z5.b, z6.b\n" + ".inst 0x45c2982b // ummla z11.s, z1.b, z2.b\n" + ".inst 0x45c29873 // ummla z19.s, z3.b, z2.b\n" + ".inst 0x45c298bb // ummla z27.s, z5.b, z2.b\n" + ".inst 0x45c0982f // ummla z15.s, z1.b, z0.b\n" + ".inst 0x45c09877 // ummla z23.s, z3.b, z0.b\n" + ".inst 0x45c098bf // ummla z31.s, z5.b, z0.b\n" "54:" // Height 5: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" "cmp x28, x20\n" "bne 49b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x24, x9, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "uzp1 z7.d, z8.d, z12.d\n" + "add x23, x9, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" + "uzp1 z2.d, z8.d, z12.d\n" "add x21, x22, x20, LSL #2\n" + "add x20, x21, x20, LSL #2\n" "uzp2 z8.d, z8.d, z12.d\n" - "uzp1 z12.d, z9.d, z13.d\n" + "uzp1 z1.d, z9.d, z13.d\n" "uzp2 z9.d, z9.d, z13.d\n" - "uzp1 z13.d, z10.d, z14.d\n" - "st1w { z7.s }, p4, [x9]\n" + "uzp1 z0.d, z10.d, z14.d\n" + "st1w { z2.s }, p4, [x9]\n" "uzp2 z10.d, z10.d, z14.d\n" - "uzp1 z14.d, z11.d, z15.d\n" - "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" + "uzp1 z2.d, z11.d, z15.d\n" + "st1w { z1.s }, p3, [x9, #1, MUL VL]\n" "uzp2 z11.d, z11.d, z15.d\n" - "uzp1 z15.d, z16.d, z20.d\n" - "st1w { z13.s }, p2, [x9, #2, MUL VL]\n" + "uzp1 z1.d, z16.d, z20.d\n" + "st1w { z0.s }, p2, [x9, #2, MUL VL]\n" "uzp2 z16.d, z16.d, z20.d\n" - "uzp1 z20.d, z17.d, z21.d\n" - "st1w { z14.s }, p1, [x9, #3, MUL VL]\n" + "uzp1 z0.d, z17.d, z21.d\n" + "st1w { z2.s }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" "uzp2 z17.d, z17.d, z21.d\n" "uzp1 z21.d, z18.d, z22.d\n" - "st1w { z8.s }, p4, [x24]\n" + "st1w { z8.s }, p4, [x23]\n" "uzp2 z18.d, z18.d, z22.d\n" - "uzp1 z22.d, z19.d, z23.d\n" - "st1w { z9.s }, p3, [x24, #1, MUL VL]\n" + "uzp1 z20.d, z19.d, z23.d\n" + "st1w { z9.s }, p3, [x23, #1, MUL VL]\n" "uzp2 z19.d, z19.d, z23.d\n" "uzp1 z24.d, z24.d, z28.d\n" - "st1w { z10.s }, p2, [x24, #2, MUL VL]\n" + "st1w { z10.s }, p2, [x23, #2, MUL VL]\n" "uzp1 z25.d, z25.d, z29.d\n" "uzp1 z26.d, z26.d, z30.d\n" - "st1w { z11.s }, p1, [x24, #3, MUL VL]\n" + "st1w { z11.s }, p1, [x23, #3, MUL VL]\n" "uzp1 z27.d, z27.d, z31.d\n" - "st1w { z15.s }, p4, [x23]\n" - "st1w { z20.s }, p3, [x23, #1, MUL VL]\n" - "st1w { z21.s }, p2, [x23, #2, MUL VL]\n" - "st1w { z22.s }, p1, [x23, #3, MUL VL]\n" - "st1w { z16.s }, p4, [x22]\n" - "st1w { z17.s }, p3, [x22, #1, MUL VL]\n" - "st1w { z18.s }, p2, [x22, #2, MUL VL]\n" - "st1w { z19.s }, p1, [x22, #3, MUL VL]\n" - "st1w { z24.s }, p4, [x21]\n" - "st1w { z25.s }, p3, [x21, #1, MUL VL]\n" - "st1w { z26.s }, p2, [x21, #2, MUL VL]\n" - "st1w { z27.s }, p1, [x21, #3, MUL VL]\n" + "st1w { z1.s }, p4, [x22]\n" + "st1w { z0.s }, p3, [x22, #1, MUL VL]\n" + "st1w { z21.s }, p2, [x22, #2, MUL VL]\n" + "st1w { z20.s }, p1, [x22, #3, MUL VL]\n" + "st1w { z16.s }, p4, [x21]\n" + "st1w { z17.s }, p3, [x21, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x21, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x21, #3, MUL VL]\n" + "st1w { z24.s }, p4, [x20]\n" + "st1w { z25.s }, p3, [x20, #1, MUL VL]\n" + "st1w { z26.s }, p2, [x20, #2, MUL VL]\n" + "st1w { z27.s }, p1, [x20, #3, MUL VL]\n" "55:" // Height 5: Writeback done "decw x11, ALL, MUL #4\n" "cmp x11, XZR\n" @@ -1307,26 +1307,26 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "add x24, x9, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" - "ld1w { z9.s }, p4/Z, [x9]\n" + "ld1w { z17.s }, p4/Z, [x9]\n" "add x22, x23, x20, LSL #2\n" "add x21, x22, x20, LSL #2\n" - "ld1w { z10.s }, p3/Z, [x9, #1, MUL VL]\n" - "ld1w { z11.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z18.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z20.s }, p2/Z, [x9, #2, MUL VL]\n" "add x20, x21, x20, LSL #2\n" "ld1w { z16.s }, p1/Z, [x9, #3, MUL VL]\n" "ld1w { z12.s }, p4/Z, [x24]\n" - "zip1 z8.d, z9.d, z12.d\n" + "zip1 z8.d, z17.d, z12.d\n" "ld1w { z13.s }, p3/Z, [x24, #1, MUL VL]\n" "ld1w { z14.s }, p2/Z, [x24, #2, MUL VL]\n" - "zip2 z12.d, z9.d, z12.d\n" - "zip1 z9.d, z10.d, z13.d\n" + "zip2 z12.d, z17.d, z12.d\n" + "zip1 z9.d, z18.d, z13.d\n" "ld1w { z15.s }, p1/Z, [x24, #3, MUL VL]\n" "ld1w { z17.s }, p4/Z, [x23]\n" - "zip2 z13.d, z10.d, z13.d\n" - "zip1 z10.d, z11.d, z14.d\n" + "zip2 z13.d, z18.d, z13.d\n" + "zip1 z10.d, z20.d, z14.d\n" "ld1w { z18.s }, p3/Z, [x23, #1, MUL VL]\n" "ld1w { z19.s }, p2/Z, [x23, #2, MUL VL]\n" - "zip2 z14.d, z11.d, z14.d\n" + "zip2 z14.d, z20.d, z14.d\n" "zip1 z11.d, z16.d, z15.d\n" "ld1w { z24.s }, p1/Z, [x23, #3, MUL VL]\n" "ld1w { z20.s }, p4/Z, [x22]\n" @@ -1344,7 +1344,7 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "ld1w { z27.s }, p2/Z, [x21, #2, MUL VL]\n" "zip2 z22.d, z19.d, z22.d\n" "zip1 z19.d, z24.d, z23.d\n" - "ld1w { z6.s }, p1/Z, [x21, #3, MUL VL]\n" + "ld1w { z0.s }, p1/Z, [x21, #3, MUL VL]\n" "ld1w { z28.s }, p4/Z, [x20]\n" "zip2 z23.d, z24.d, z23.d\n" "zip1 z24.d, z25.d, z28.d\n" @@ -1356,8 +1356,8 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "zip2 z29.d, z26.d, z29.d\n" "zip1 z26.d, z27.d, z30.d\n" "zip2 z30.d, z27.d, z30.d\n" - "zip1 z27.d, z6.d, z31.d\n" - "zip2 z31.d, z6.d, z31.d\n" + "zip1 z27.d, z0.d, z31.d\n" + "zip2 z31.d, z0.d, z31.d\n" "b 59f\n" "58:" // Height 6: no accumulate "mov z8.s, #0x0\n" @@ -1389,16 +1389,16 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "60:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 61f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "ldr x21, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "ldr x21, [x20, #0x28]\n" "cbnz x28, 62f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -1410,184 +1410,184 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "b 62f\n" "61:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" - "add x22, x23, x20\n" - "add x21, x22, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" + "add x23, x24, x21\n" + "add x22, x23, x21\n" + "add x21, x22, x21\n" "62:" // Height 6: input setup done "cmp x27, #0x10\n" "ble 64f\n" "63:" // Height 6: Multiply loop: Main loop head "whilelt p0.b, XZR, x27\n" - "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" - "ld1rqb { z3.b }, p0/Z, [x24]\n" - "ld1rqb { z4.b }, p0/Z, [x23]\n" - "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z4.d\n" - "ld1rqb { z5.b }, p0/Z, [x22]\n" - "ld1rqb { z6.b }, p0/Z, [x21]\n" - "trn2 z3.d, z3.d, z4.d\n" - "trn1 z4.d, z5.d, z6.d\n" - "trn2 z5.d, z5.d, z6.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c79850 // ummla z16.s, z2.b, z7.b\n" - ".inst 0x45c79898 // ummla z24.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1rqb { z7.b }, p0/Z, [x26]\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" + "trn1 z6.d, z7.d, z0.d\n" + "ld1rqb { z5.b }, p0/Z, [x24]\n" + "ld1rqb { z1.b }, p0/Z, [x23]\n" + "trn2 z7.d, z7.d, z0.d\n" + "trn1 z4.d, z5.d, z1.d\n" + "ld1rqb { z3.b }, p0/Z, [x22]\n" + "ld1rqb { z0.b }, p0/Z, [x21]\n" + "trn2 z5.d, z5.d, z1.d\n" + "trn1 z2.d, z3.d, z0.d\n" + "trn2 z3.d, z3.d, z0.d\n" + "ld1b { z1.b }, p5/Z, [x10]\n" + "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45c198c8 // ummla z8.s, z6.b, z1.b\n" + ".inst 0x45c19890 // ummla z16.s, z4.b, z1.b\n" + ".inst 0x45c19858 // ummla z24.s, z2.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #2, MUL VL]\n" "sub x27, x27, #0x10\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - ".inst 0x45c69854 // ummla z20.s, z2.b, z6.b\n" + ".inst 0x45c098cc // ummla z12.s, z6.b, z0.b\n" + ".inst 0x45c09894 // ummla z20.s, z4.b, z0.b\n" "cmp x27, #0x10\n" "add x26, x26, #0x10\n" - ".inst 0x45c6989c // ummla z28.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" + ".inst 0x45c0985c // ummla z28.s, z2.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45c198c9 // ummla z9.s, z6.b, z1.b\n" "add x25, x25, #0x10\n" - ".inst 0x45c79851 // ummla z17.s, z2.b, z7.b\n" - ".inst 0x45c79899 // ummla z25.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x45c19891 // ummla z17.s, z4.b, z1.b\n" + ".inst 0x45c19859 // ummla z25.s, z2.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #4, MUL VL]\n" "add x24, x24, #0x10\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - ".inst 0x45c69855 // ummla z21.s, z2.b, z6.b\n" + ".inst 0x45c098cd // ummla z13.s, z6.b, z0.b\n" + ".inst 0x45c09895 // ummla z21.s, z4.b, z0.b\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - ".inst 0x45c6989d // ummla z29.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" + ".inst 0x45c0985d // ummla z29.s, z2.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45c198ca // ummla z10.s, z6.b, z1.b\n" "add x21, x21, #0x10\n" - ".inst 0x45c79852 // ummla z18.s, z2.b, z7.b\n" - ".inst 0x45c7989a // ummla z26.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - ".inst 0x45c69856 // ummla z22.s, z2.b, z6.b\n" - ".inst 0x45c6989e // ummla z30.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45c19892 // ummla z18.s, z4.b, z1.b\n" + ".inst 0x45c1985a // ummla z26.s, z2.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45c098ce // ummla z14.s, z6.b, z0.b\n" + ".inst 0x45c09896 // ummla z22.s, z4.b, z0.b\n" + ".inst 0x45c0985e // ummla z30.s, z2.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #16\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" - ".inst 0x45c79853 // ummla z19.s, z2.b, z7.b\n" - ".inst 0x45c7989b // ummla z27.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-8, MUL VL]\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - ".inst 0x45c69857 // ummla z23.s, z2.b, z6.b\n" - ".inst 0x45c6989f // ummla z31.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-7, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c79870 // ummla z16.s, z3.b, z7.b\n" - ".inst 0x45c798b8 // ummla z24.s, z5.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-6, MUL VL]\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - ".inst 0x45c69874 // ummla z20.s, z3.b, z6.b\n" - ".inst 0x45c698bc // ummla z28.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-5, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c79871 // ummla z17.s, z3.b, z7.b\n" - ".inst 0x45c798b9 // ummla z25.s, z5.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-4, MUL VL]\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - ".inst 0x45c69875 // ummla z21.s, z3.b, z6.b\n" - ".inst 0x45c698bd // ummla z29.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-3, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c79872 // ummla z18.s, z3.b, z7.b\n" - ".inst 0x45c798ba // ummla z26.s, z5.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #-2, MUL VL]\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - ".inst 0x45c69876 // ummla z22.s, z3.b, z6.b\n" - ".inst 0x45c698be // ummla z30.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #-1, MUL VL]\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c79873 // ummla z19.s, z3.b, z7.b\n" - ".inst 0x45c798bb // ummla z27.s, z5.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" - ".inst 0x45c69877 // ummla z23.s, z3.b, z6.b\n" - ".inst 0x45c698bf // ummla z31.s, z5.b, z6.b\n" + ".inst 0x45c198cb // ummla z11.s, z6.b, z1.b\n" + ".inst 0x45c19893 // ummla z19.s, z4.b, z1.b\n" + ".inst 0x45c1985b // ummla z27.s, z2.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #-8, MUL VL]\n" + ".inst 0x45c098cf // ummla z15.s, z6.b, z0.b\n" + ".inst 0x45c09897 // ummla z23.s, z4.b, z0.b\n" + ".inst 0x45c0985f // ummla z31.s, z2.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #-7, MUL VL]\n" + ".inst 0x45c198e8 // ummla z8.s, z7.b, z1.b\n" + ".inst 0x45c198b0 // ummla z16.s, z5.b, z1.b\n" + ".inst 0x45c19878 // ummla z24.s, z3.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #-6, MUL VL]\n" + ".inst 0x45c098ec // ummla z12.s, z7.b, z0.b\n" + ".inst 0x45c098b4 // ummla z20.s, z5.b, z0.b\n" + ".inst 0x45c0987c // ummla z28.s, z3.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #-5, MUL VL]\n" + ".inst 0x45c198e9 // ummla z9.s, z7.b, z1.b\n" + ".inst 0x45c198b1 // ummla z17.s, z5.b, z1.b\n" + ".inst 0x45c19879 // ummla z25.s, z3.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #-4, MUL VL]\n" + ".inst 0x45c098ed // ummla z13.s, z7.b, z0.b\n" + ".inst 0x45c098b5 // ummla z21.s, z5.b, z0.b\n" + ".inst 0x45c0987d // ummla z29.s, z3.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #-3, MUL VL]\n" + ".inst 0x45c198ea // ummla z10.s, z7.b, z1.b\n" + ".inst 0x45c198b2 // ummla z18.s, z5.b, z1.b\n" + ".inst 0x45c1987a // ummla z26.s, z3.b, z1.b\n" + "ld1b { z1.b }, p5/Z, [x10, #-2, MUL VL]\n" + ".inst 0x45c098ee // ummla z14.s, z7.b, z0.b\n" + ".inst 0x45c098b6 // ummla z22.s, z5.b, z0.b\n" + ".inst 0x45c0987e // ummla z30.s, z3.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #-1, MUL VL]\n" + ".inst 0x45c198eb // ummla z11.s, z7.b, z1.b\n" + ".inst 0x45c198b3 // ummla z19.s, z5.b, z1.b\n" + ".inst 0x45c1987b // ummla z27.s, z3.b, z1.b\n" + ".inst 0x45c098ef // ummla z15.s, z7.b, z0.b\n" + ".inst 0x45c098b7 // ummla z23.s, z5.b, z0.b\n" + ".inst 0x45c0987f // ummla z31.s, z3.b, z0.b\n" "bgt 63b\n" "64:" // Height 6: Multiply loop: Single iteration only "whilelt p0.b, XZR, x27\n" "ld1rqb { z1.b }, p0/Z, [x26]\n" - "ld1rqb { z2.b }, p0/Z, [x25]\n" - "trn1 z0.d, z1.d, z2.d\n" + "ld1rqb { z0.b }, p0/Z, [x25]\n" + "trn1 z7.d, z1.d, z0.d\n" "ld1rqb { z3.b }, p0/Z, [x24]\n" - "ld1rqb { z4.b }, p0/Z, [x23]\n" - "trn2 z1.d, z1.d, z2.d\n" - "trn1 z2.d, z3.d, z4.d\n" + "ld1rqb { z2.b }, p0/Z, [x23]\n" + "trn2 z1.d, z1.d, z0.d\n" + "trn1 z6.d, z3.d, z2.d\n" "ld1rqb { z5.b }, p0/Z, [x22]\n" - "ld1rqb { z6.b }, p0/Z, [x21]\n" - "trn2 z3.d, z3.d, z4.d\n" - "trn1 z4.d, z5.d, z6.d\n" - "trn2 z5.d, z5.d, z6.d\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79808 // ummla z8.s, z0.b, z7.b\n" - ".inst 0x45c79850 // ummla z16.s, z2.b, z7.b\n" - ".inst 0x45c79898 // ummla z24.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" + "ld1rqb { z0.b }, p0/Z, [x21]\n" + "trn2 z3.d, z3.d, z2.d\n" + "trn1 z4.d, z5.d, z0.d\n" + "trn2 z5.d, z5.d, z0.d\n" + "ld1b { z2.b }, p5/Z, [x10]\n" + "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45c298e8 // ummla z8.s, z7.b, z2.b\n" + ".inst 0x45c298d0 // ummla z16.s, z6.b, z2.b\n" + ".inst 0x45c29898 // ummla z24.s, z4.b, z2.b\n" + "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n" "subs x27, x27, #0x8\n" - ".inst 0x45c6980c // ummla z12.s, z0.b, z6.b\n" - ".inst 0x45c69854 // ummla z20.s, z2.b, z6.b\n" - ".inst 0x45c6989c // ummla z28.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79809 // ummla z9.s, z0.b, z7.b\n" - ".inst 0x45c79851 // ummla z17.s, z2.b, z7.b\n" - ".inst 0x45c79899 // ummla z25.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x45c6980d // ummla z13.s, z0.b, z6.b\n" - ".inst 0x45c69855 // ummla z21.s, z2.b, z6.b\n" - ".inst 0x45c6989d // ummla z29.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7980a // ummla z10.s, z0.b, z7.b\n" - ".inst 0x45c79852 // ummla z18.s, z2.b, z7.b\n" - ".inst 0x45c7989a // ummla z26.s, z4.b, z7.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c6980e // ummla z14.s, z0.b, z6.b\n" - ".inst 0x45c69856 // ummla z22.s, z2.b, z6.b\n" - ".inst 0x45c6989e // ummla z30.s, z4.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" - ".inst 0x45c7980b // ummla z11.s, z0.b, z7.b\n" + ".inst 0x45c098ec // ummla z12.s, z7.b, z0.b\n" + ".inst 0x45c098d4 // ummla z20.s, z6.b, z0.b\n" + ".inst 0x45c0989c // ummla z28.s, z4.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45c298e9 // ummla z9.s, z7.b, z2.b\n" + ".inst 0x45c298d1 // ummla z17.s, z6.b, z2.b\n" + ".inst 0x45c29899 // ummla z25.s, z4.b, z2.b\n" + "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x45c098ed // ummla z13.s, z7.b, z0.b\n" + ".inst 0x45c098d5 // ummla z21.s, z6.b, z0.b\n" + ".inst 0x45c0989d // ummla z29.s, z4.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45c298ea // ummla z10.s, z7.b, z2.b\n" + ".inst 0x45c298d2 // ummla z18.s, z6.b, z2.b\n" + ".inst 0x45c2989a // ummla z26.s, z4.b, z2.b\n" + "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45c098ee // ummla z14.s, z7.b, z0.b\n" + ".inst 0x45c098d6 // ummla z22.s, z6.b, z0.b\n" + ".inst 0x45c0989e // ummla z30.s, z4.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n" + ".inst 0x45c298eb // ummla z11.s, z7.b, z2.b\n" "addvl x10, x10, #8\n" - ".inst 0x45c79853 // ummla z19.s, z2.b, z7.b\n" - ".inst 0x45c7989b // ummla z27.s, z4.b, z7.b\n" - ".inst 0x45c6980f // ummla z15.s, z0.b, z6.b\n" - ".inst 0x45c69857 // ummla z23.s, z2.b, z6.b\n" - ".inst 0x45c6989f // ummla z31.s, z4.b, z6.b\n" + ".inst 0x45c298d3 // ummla z19.s, z6.b, z2.b\n" + ".inst 0x45c2989b // ummla z27.s, z4.b, z2.b\n" + ".inst 0x45c098ef // ummla z15.s, z7.b, z0.b\n" + ".inst 0x45c098d7 // ummla z23.s, z6.b, z0.b\n" + ".inst 0x45c0989f // ummla z31.s, z4.b, z0.b\n" "ble 65f\n" - "ld1b { z7.b }, p5/Z, [x10]\n" - "ld1b { z6.b }, p5/Z, [x10, #1, MUL VL]\n" - ".inst 0x45c79828 // ummla z8.s, z1.b, z7.b\n" - ".inst 0x45c79870 // ummla z16.s, z3.b, z7.b\n" - ".inst 0x45c798b8 // ummla z24.s, z5.b, z7.b\n" - ".inst 0x45c6982c // ummla z12.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #2, MUL VL]\n" - ".inst 0x45c69874 // ummla z20.s, z3.b, z6.b\n" - ".inst 0x45c698bc // ummla z28.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #3, MUL VL]\n" - ".inst 0x45c79829 // ummla z9.s, z1.b, z7.b\n" - ".inst 0x45c79871 // ummla z17.s, z3.b, z7.b\n" - ".inst 0x45c798b9 // ummla z25.s, z5.b, z7.b\n" - ".inst 0x45c6982d // ummla z13.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #4, MUL VL]\n" - ".inst 0x45c69875 // ummla z21.s, z3.b, z6.b\n" - ".inst 0x45c698bd // ummla z29.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #5, MUL VL]\n" - ".inst 0x45c7982a // ummla z10.s, z1.b, z7.b\n" - ".inst 0x45c79872 // ummla z18.s, z3.b, z7.b\n" - ".inst 0x45c798ba // ummla z26.s, z5.b, z7.b\n" - ".inst 0x45c6982e // ummla z14.s, z1.b, z6.b\n" - "ld1b { z7.b }, p5/Z, [x10, #6, MUL VL]\n" - ".inst 0x45c69876 // ummla z22.s, z3.b, z6.b\n" - ".inst 0x45c698be // ummla z30.s, z5.b, z6.b\n" - "ld1b { z6.b }, p5/Z, [x10, #7, MUL VL]\n" + "ld1b { z2.b }, p5/Z, [x10]\n" + "ld1b { z0.b }, p5/Z, [x10, #1, MUL VL]\n" + ".inst 0x45c29828 // ummla z8.s, z1.b, z2.b\n" + ".inst 0x45c29870 // ummla z16.s, z3.b, z2.b\n" + ".inst 0x45c298b8 // ummla z24.s, z5.b, z2.b\n" + ".inst 0x45c0982c // ummla z12.s, z1.b, z0.b\n" + "ld1b { z2.b }, p5/Z, [x10, #2, MUL VL]\n" + ".inst 0x45c09874 // ummla z20.s, z3.b, z0.b\n" + ".inst 0x45c098bc // ummla z28.s, z5.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #3, MUL VL]\n" + ".inst 0x45c29829 // ummla z9.s, z1.b, z2.b\n" + ".inst 0x45c29871 // ummla z17.s, z3.b, z2.b\n" + ".inst 0x45c298b9 // ummla z25.s, z5.b, z2.b\n" + ".inst 0x45c0982d // ummla z13.s, z1.b, z0.b\n" + "ld1b { z2.b }, p5/Z, [x10, #4, MUL VL]\n" + ".inst 0x45c09875 // ummla z21.s, z3.b, z0.b\n" + ".inst 0x45c098bd // ummla z29.s, z5.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #5, MUL VL]\n" + ".inst 0x45c2982a // ummla z10.s, z1.b, z2.b\n" + ".inst 0x45c29872 // ummla z18.s, z3.b, z2.b\n" + ".inst 0x45c298ba // ummla z26.s, z5.b, z2.b\n" + ".inst 0x45c0982e // ummla z14.s, z1.b, z0.b\n" + "ld1b { z2.b }, p5/Z, [x10, #6, MUL VL]\n" + ".inst 0x45c09876 // ummla z22.s, z3.b, z0.b\n" + ".inst 0x45c098be // ummla z30.s, z5.b, z0.b\n" + "ld1b { z0.b }, p5/Z, [x10, #7, MUL VL]\n" "addvl x10, x10, #8\n" - ".inst 0x45c7982b // ummla z11.s, z1.b, z7.b\n" - ".inst 0x45c79873 // ummla z19.s, z3.b, z7.b\n" - ".inst 0x45c798bb // ummla z27.s, z5.b, z7.b\n" - ".inst 0x45c6982f // ummla z15.s, z1.b, z6.b\n" - ".inst 0x45c69877 // ummla z23.s, z3.b, z6.b\n" - ".inst 0x45c698bf // ummla z31.s, z5.b, z6.b\n" + ".inst 0x45c2982b // ummla z11.s, z1.b, z2.b\n" + ".inst 0x45c29873 // ummla z19.s, z3.b, z2.b\n" + ".inst 0x45c298bb // ummla z27.s, z5.b, z2.b\n" + ".inst 0x45c0982f // ummla z15.s, z1.b, z0.b\n" + ".inst 0x45c09877 // ummla z23.s, z3.b, z0.b\n" + ".inst 0x45c098bf // ummla z31.s, z5.b, z0.b\n" "65:" // Height 6: Multiply loop: multiply skip "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1596,7 +1596,7 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "add x24, x9, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" - "uzp1 z7.d, z8.d, z12.d\n" + "uzp1 z0.d, z8.d, z12.d\n" "add x22, x23, x20, LSL #2\n" "add x21, x22, x20, LSL #2\n" "uzp2 z8.d, z8.d, z12.d\n" @@ -1604,7 +1604,7 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "add x20, x21, x20, LSL #2\n" "uzp2 z9.d, z9.d, z13.d\n" "uzp1 z13.d, z10.d, z14.d\n" - "st1w { z7.s }, p4, [x9]\n" + "st1w { z0.s }, p4, [x9]\n" "uzp2 z10.d, z10.d, z14.d\n" "uzp1 z14.d, z11.d, z15.d\n" "st1w { z12.s }, p3, [x9, #1, MUL VL]\n" @@ -1664,7 +1664,6 @@ void sve_hybrid_u8u32_mmla_6x4VL ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "68:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" @@ -1672,4 +1671,4 @@ void sve_hybrid_u8u32_mmla_6x4VL ( } } // namespace arm_gemm -#endif // __ARM_FEATURE_SVE +#endif // ARM_COMPUTE_ENABLE_SVE -- cgit v1.2.1